I explored dataset splitting and json files and found that although support images used for validation are removed from json with validation instances, those images are not quite removed from train subset. The annotations with categories from validation are removed, but if there is another label in the image that is used for training the image are remained.
For example, the image:
{'license': 3, 'file_name': '000000386164.jpg', 'coco_url': 'http://images.cocodataset.org/train2017/000000386164.jpg', 'height': 640, 'width': 480, 'date_captured': '2013-11-14 21:06:13', 'flickr_url': 'http://farm6.staticflickr.com/5216/5384496530_f0aa6628a2_z.jpg', 'id': 386164}
and it's annotations with different labels are presented both in train json and final_split_voc_10_shot_instances_train2017.
Although categories are differs and there is no direct overfitting, I think there is a little possibility that something could went wrong.
Hello! Thanks very much for your work.
I explored dataset splitting and json files and found that although support images used for validation are removed from json with validation instances, those images are not quite removed from train subset. The annotations with categories from validation are removed, but if there is another label in the image that is used for training the image are remained.
For example, the image: {'license': 3, 'file_name': '000000386164.jpg', 'coco_url': 'http://images.cocodataset.org/train2017/000000386164.jpg', 'height': 640, 'width': 480, 'date_captured': '2013-11-14 21:06:13', 'flickr_url': 'http://farm6.staticflickr.com/5216/5384496530_f0aa6628a2_z.jpg', 'id': 386164} and it's annotations with different labels are presented both in train json and final_split_voc_10_shot_instances_train2017. Although categories are differs and there is no direct overfitting, I think there is a little possibility that something could went wrong.
Prease correct me if I am wrong! :)